AI for Scientific Discovery

AlphaFold, AlphaProof, GNoME. AI is now a tool in scientific workflows, not just demos. Here is what works and what the next frontier looks like.

What AI has actually delivered

• AlphaFold 2/3: protein structure prediction, transformed structural biology.
• GNoME: 2.2 million predicted stable crystal structures, verified for hundreds of new materials.
• AlphaProof / AlphaGeometry: silver-medal IMO performance on formal proofs and geometry.
• AlphaMissense: classification of protein variants for clinical relevance.

The architecture pattern

The successful systems combine learned models with explicit search. Neural networks evaluate candidates; search algorithms (Monte Carlo, beam search, formal proof search) explore the candidate space. Pure end-to-end transformers haven’t been the winning recipe in scientific discovery so far.

Limits

Domain-specific. Each system was built around a specific scientific problem with verifiable outcomes. General-purpose “AI scientist” systems have produced papers and citations, but not yet new findings of comparable importance.

Next frontier

Drug discovery (molecule-level), climate modelling (downscaling), fluid dynamics (turbulence simulation), and theorem proving in pure mathematics. All have similar structure: large search space + verifiable evaluation + huge economic value if it works.